MDB-Net: A dual-stream synergistic adaptation network with foundation model-guided semantic alignment for brain tumor segmentation

Precise brain tumor segmentation requires balancing high-capacity modeling with the scarcity of annotated medical data, which we refer to as the Capacity–Data Paradox. Although foundation models provide rich anatomical semantics, directly deploying them for dense medical segmentation may introduce considerable computational overhead and may not sufficiently preserve fine-grained boundary details. To address this issue, we propose MDB-Net, a Dual-Stream Synergistic Adaptation Framework that balances representational capacity and inference efficiency. MDB-Net integrates a Dual-Stream Hybrid Encoding Block (DS-HEB) with Spatial-Channel Fusion to jointly capture local textures and long-range dependencies. To enhance deep feature reasoning, a Sparse Reasoning Module (SRM) transforms passive skip connections into active reasoning units, improving boundary-aware representation. Furthermore, we introduce a Foundation Model-guided Semantic Alignment (FMSA) strategy, where a frozen MedSAM encoder provides training-time anatomical priors through feature-level alignment. This design encourages structural consistency without adding inference-time overhead. Extensive experiments on BraTS 2021 and BraTS 2024 demonstrate that MDB-Net achieves strong segmentation performance, with Mean DSC values of 88.50% and 93.04%, respectively, showing favorable generalization compared with recent CNN-, Transformer-, MLP-, and Mamba-based segmentation methods.